Second day of Needham's 64th Annual Growth Conference. I'm Quinn Bolton. I cover the semiconductor industry for Needham. It's my pleasure to host this fireside chat with Ambarella. The company's a leading supplier of low-power, deep neural network AI vision processors for edge applications, including automobile security cameras and robotics. Ambarella's low-power systems on chip offer high-resolution video compression, advanced image and radar processing, and powerful deep neural network processing to enable intelligent perception, fusion, and planning. Joining me on stage is President and CEO, Fermi Wang. We also have John Young, VP of Finance, and soon to be CFO, I think, just within the next month or so, and Louis Gerhardy , VP of Corporate Development, in the audience. And so, Fermi, thank you for joining us.
Thank you, Quinn.
I wanted to start off, obviously, last week was a big week with CES. You had a number of announcements at the show, the N1, a new development platform, two new members of CV3, and so maybe touch on the highlights of CES as you saw it.
Right. So I think first thing is that, we have been working in this automotive roadmap for a while, and we basically announced two more chips, 655, 635. Basically, this really providing with our previous announcement of 685, we provide a complete roadmap on the CV3 for a high-end Level 3, Level 4 car, all the way down to the city level, a highway level of autonomous driving. So this is really give a different performance and price point for our customer to choose, and more importantly, since software stack apply to all three of them. And when we say same software stack, really is not rewriting, it's really just recompile the code, you are working on different chips.
And that gives a customer flexibility to move from a low-end chip to a high-end chip, based on their products. That's the first announcement. The second important announcement is that we announced our software stack, complete software stack, that is 100% owned by Ambarella technology, but also is AI-enabled. And we demo that software stack running on CV3 with a Level 3 autonomous car at the CES. That's, I think, is important, very important milestone for automotive application. Another important announcement we did is we announced N1, which is our first Gen AI inference engine. Now, we talked about this two quarters ago, and we talked to a lot of customers, we sampled to customers, and we're ready to take that into production.
The silicon is available. We demo multiple Llama-based model at the CES. One of them is very interesting. We feed a live video camera to a Llama model, and the Llama start describing what they see on the video, right? So this is not, y ou know, in the past, when a AI camera need to identify objects, you need to train the model to understand this is a man, this is a car, this is a license plate. The Llama model did not only that, but also describing how many people there, you know, you know, what kind of color car, the jacket, who are wearing ties. Things that people they start summarize. Just like, you know, how the ChatGPT summarize things for you.
So you can imagine that we're using Llama for the video application in the future. So that's just one example, but we believe there are many different way. And we also continue to believe that each inference will be a huge market, and we are definitely one of the leaders in the market right now. So that's third.
The fourth one, which is less impressive for the investor, but is very important for engineering. Basically, we announce a new Cooper Developer Platform , which is a platform that will apply to all of our current silicon and give our customer and software developer a much easier platform that integrate not only just the low-level drivers, you know, our mid-level applications, but more importantly, for example, all our AI software, AI compiler that we generate to help our any AI modeling developers. They can quickly convert the model to run our chip. Everything we integrate into one platform, and it's very important development tools for us. So that's the highlights for CES.
That's a lot. So I've got a few follow-up questions. On the Cooper Developer Platform , you mentioned the compiler.
Right.
Does it also have sort of that, that CUDA equivalent, low-level, low sort of level software that helps, you know, place the, the parts of the model within the CVflow architecture to help optimize that process, and so it kind of hides that complexity?
Absolutely. Although today we are focusing on Llama-based model, but if you run Llama, you can easily, using this tool, Cooper software development platform, to compile your model, and that also for will try to help you to allocate resources, and then run on our silicon automatically. I think that's the goal, and definitely we are in the current release to our customer, the Llama support is there.
Okay. Last week, when we saw Louis, we talked about the new software stack. I think every module now is deep learning-based. Talk about the importance of that, and then also, what was the reception from customers to the new software stack?
Right. So first of all, it's all AI-based. You know, people talking about, you know, end-to-end AI software stack for the autonomous driving, and we cannot claim that just yet, but however, every module, it's not just perception. In the past, people thought, "Oh, a neural network is only important for perception," but we changed that opinion. We think that even for the path planning to the high-level control, all of them can be done AI, and we did that, and it's being implemented on our software stack. And I think not many software developers can show that, and I think we get a huge feedback on that one, because I think we're not only showing the performance that we can achieve, but also it's easier to add features. Just give you a quick example.
The first time I sat in this car, you know, you know, we, with our teams in Italy, in Europe, you have a lot of roundabouts, right? In the past, when we, when our autonomous driving software approached the roundabout, it stop before the circle and start looking around. The first time we used this AI software, I was in there, and we found out that the car didn't stop at the, the line, and in fact, it moved the car over the line, just like normal drivers, right? And then because it provide better visibility of the cameras, then I ask, "Oh, Gee, how does the car do that?" He says, "Learn from our, you know, our training videos.
Videos, yeah.
So that just shows you that this kind of AI approach will gradually help people easier to adopt how people drive, not just using, you know, rule-based assumption, like what we did before. So I start seeing a lot of potential on this AI-based software, and we're going to continue to upgrade on that.
You mentioned the extension of the CV3 family, the 635 and the 655 family members. Talk about the price points of those solutions. I think your 685, the high end of the family, can sell for up to $400.
Right.
What's a more typical price point for some of these newer?
For 635, we're targeting really, you know, let's call it equivalent to Tesla's Autopilot. It's $100 ASP, plus or minus. On the 655, is $200+, $200. So you can see that on the low end, 635, $100, $200, 655, $400, of a 685. Just a round number for this, give you, give our customer a sense of where our price and performance is going to be.
Great. Another sort of announcement last week was you won a CES Innovation Award for the Oculii centralized radar approach, and so maybe spend a minute. Some in the room may be familiar with Oculii, but talk about the benefits of centralized radar, and what kind of traction are you seeing with customers?
Right.
I know you have partnerships with some of the big radar-
Right
Solutions providers.
Right. So I think, you know, today, nobody's arguing whether you need a domain controller for cameras anymore. You remember, I would say two years ago, people still arguing on that. Now, nobody arguing on that. But people still saying, you know, whether radar should be centralized or not. I think this is, in my mind, it is a foregone conclusion already. Because all of the signal processing on the autonomous driving car should be centralized, because there's a huge benefit on that. And we are the first one to really showing a centralized radar solution because not only we have the algorithm, but also we have silicon architectures to support it, and we show the benefit of that. So I think the integration of a camera and radar processing is there already.
People see it, so now the only question is whether you want to continue to use satellite or you use edge radar, which, you know, has been used for 20 years, or you want to use centralized radar, which give you much better integration, quality, and the performance. That, but however, you have to really give up the current architecture. I think convincing our current customer, OEM or Tier 1 , to say, "Okay, you need to go to the new concept," that's difficult. But however, we are showing enough values and advantages so that people can take advantage of the centralized radar.
Great. Last question for you, just on CES, you had examples of AI moving into sort of non-auto, non, you know, kind of security applications. Insta360 was showing sort of CV5-based action camera.
Right.
You were showing some demos with AI-enhanced video conferencing. Talk about a couple of those opportunities, and do you think that may give some longevity to the more legacy part of your business?
You know, I think starting a few years ago, we start talking about, in the past, our application focus on video. And, in the past few years, we think is video plus edge AI. So any cameras who want to use AI and want to demand more and more AI, is application that we think is suitable for us. And to our surprise, that some of those really old application that we use to address, start taking advantage of that, right? In fact, Insta360 is a great example. They use our AI engine to do much better signal processing, as well as more AI functions in their application level. So I expect that a similar AI function technology will get into more different applications, in addition to our current security cameras and the automotive.
I think that as long as people are paying attention to AI and trying to leverage our AI performance into a product line, that will be application we can help.
Perfect. Wanted to switch to some just more general, you know, business condition type questions. Obviously, the entire industry, but, but your business as well, affected by inventory digestion over the past year. Where do you sort of think we are in this inventory digestion process, both for the IoT business as well as your automotive business?
Right. Let's talk about IoT, because that's probably the biggest inventory problem we deal with last year. In that particular market, I think we start seeing significant booking strength that we talked about in the last quarter. I think booking improved dramatically. And that's just, that's the indication we're looking for, that we're coming out of this inventory correction problem. I would say that we are starting strong booking in Q4, and that give us indication maybe because any booking coming in Q4 probably will ship out in Q2. So I think that in the next two quarter, I think big portion of our customer will come out the this inventory problem. But not all of them will come out at the same time.
So we don't expect, you know, a V-shaped type of recovery, but we definitely feel comfortable with current street numbers that our analysts putting out, and I think that's probably the number where I feel comfortable right now.
Got it. But it seems like that order and booking momentum—
Mm-hmm.
That you saw in the October quarter has kind of continued here into the January.
Correct.
Good. Okay. Another question we get from investors, sort of your China exposure. I think there may be some misperceptions out there that you're more China-heavy or focused than you are today. Certainly, you had a bigger presence in China a number of years ago—
That's right.
With security customers. But is the China security business, is that effectively zero, and how much of the rest of the business is sort of consumed in China?
You know, the full product consumed in China, that's probably less than 15%, majority is automotive. Zero security camera. We still have some traditional camera business there. So it's. I, I would say the exposure is much less than before, like you said. And I don't expect that the percentage change a lot, right? Because when we scale up our automotive business in China, I expected that we're going to scale up our automotive, automotive business outside China also. So I don't think that 15% will change dramatically in the next few years.
Okay, perfect. We'll, we'll get to some of the opportunities in the China market for autos here in the next section.
Right.
But, lastly, on just general market questions, I think your lead times reached about 48 weeks at the peak of the COVID supply constraints. I think they have been normalizing since. Are you back to kind of what you would consider normal lead times, or do you still see, you know, do you still need to make some progress—
Sure.
Further reducing lead times?
For the lead times really depends on the process nodes, right? And in fact, I think that for 5 nanometers, although we dropped to 36 weeks, I think still there's room to be improved. For 14 and 7 and 10 nanometers, I think you are gradually back to normal.
Okay, perfect. I wanted to move now to the automotive market. On the October call, you updated your six-year automotive pipeline. It increased only 4% year-on-year to $2.4 billion from $2.3 billion, and I, I think, you know, that was seen, you know, somewhat of a disappointment by investors. I'm sure internally it was probably a disappointment. Can you talk about some of the puts and takes you saw in the pipeline? And then I'll have some follow-ups.
I think we talked about there are two major. There's some upside, there's one downside. I talk about downside first. The biggest downside is two things. One is a lot of design wins we were trying to bid on got pushed out because people making decision later or the project got pushed out. So that's one big part of downside. The other downside we talk about is really the full project we won, either that, you know, the volume got reduced or even the project that didn't go into production at the right time. So in fact, I will say that that is more of a inventory problem, and we will talk about in November timeframe. So, yeah, that's the basic downside. But also on the upside, we talk about the momentum is on the CV3 side.
In fact, we are bidding on multiple CV3 design wins, and also, you know, we definitely factoring some of the CV3 wins that we announced last year. And so those things will get into the positive side as well as more pipeline design potential design wins.
You know, kind of how do you see, you know, as you look forward to the next pipeline update, you know, what are the biggest things you think you can do to accelerate that pipeline and get that, you know, pipeline growing again?
So the pipeline, you know, that's, in fact, it's a factor in our probability to win, right? So if we're bidding on products, if we think it's a 50% chance to win, we multiply the potential size of business by 0.5. That's how we calculate our pipeline. So just increase that point five, the 50% to 100%, you double your size.
Right.
I think that getting design win solve that problem.
Okay. How important are things like your full stack software or, you know, new versions of CV3, you know, or how important are those to, you know, trying to grow that pipeline?
Both are very important because for our customer, they don't want to design a point product. They want to design, leverage, reuse a software, because everybody understand in this autonomous driving market, the biggest investment on software side. So nobody want to say, "I de-define, I design a product," and when, when for the, doesn't matter if it's high-end, middle-end, low-end, "When I move to different type of category, I need to rewrite software." So that we provide one software, one hardware portfolio, 635, 655, 685, and software by recompile, you can go to a different chip. That's hugely help our customer to minimize the investment on software side. So, and with our software stack, basically providing our OEM customer an option that either.
You know, most of our OEM want to control their own software, but our, our software can be viewed as a backup or some module of those, our software can be moving to our OEM customer software and help them to speed up the development. So that's how the customer viewing our hardware and software offering at this point.
Got it. Is software yet in the pipeline, or is that additive to the pipeline?
Software is additive because we haven't putting the software inside our pipeline yet, and of, and also because that's kind of far out, because it's all bound up with CV3. So today, in the pipeline is a CV3 hardware mainly.
Okay. Can you update us on the partnership with Continental? I think at CES, there was a press release with a quote from the head of the mobile autonomous mobility business of Continental, talking about start of production dates in 2027.
Right.
But, you know, any updates you can share with Continental?
Yeah, we have worked closely with Continental for both their ECU hardware developing further. We demoed our ECU at our booth this year. And also on the software side, we have a strong collaboration and joint development on the software stack. That's going to be inside the 2027 production. And you can imagine that we have to work in the next two years, not only just deliver function, but we also need to auto-grade and production-worthy software. And that will help our software development also. So I think we're using that software to the development to drive the maturity of our software. So the Conti development is continued to be very important for us, and I think that it's not just for that one particular project.
In fact, our ambition, Conti and us, is using this to drive Level 2 Plus consumer design wins, too.
Great. CV72 is a more China-specific member of the CV3 family. We know China has much faster design cycles than the Western OEMs. Talk about your, you know, the traction you're starting to see in China, and is CV72 a meaningful part of the pipeline today, or is it more of an opportunity?
So CV72, we talk about that, we are bidding on project in production at the end of 2025, production ramping up 2026. That's how fast, you know, China moves, right? So, and, the target is really low-end Level 2 Plus. Basically, it's autonomous driving. Sorry, the smart camera plus smart parking. So it's a 5-6 camera solution, plus radar. And the CV72 and the price range, we talk about $40. So you can imagine that really, CV72 is low-end CV3 family for $40, and we have solution all the way to $400. I give you a range of the roadmap. And, we are at the CES, we have multiple OEMs, Chinese OEM and Chinese Tier 1s, that only the ECU and the software running our car.
I think, you know, some people here might have seen it, and we are using that to get design wins with Chinese OEM. We're still targeting that, hopefully, if they stay scheduled, the end of 2025, early 2025, the sequence start seeing ramping from CVs, from CV72.
Got it, and so that, that's part of that CV3—
Yes
Ramp that you've, I think, historically said is sort of a 2026.
Right.
Yeah.
However, ASP is lower, and also, I have to say that although it's faster to design Chinese design, but also because they spin the product line so quickly, it's not a six-year project.
Right.
It's, like, more like two to three year project. So that's the downside of working with Chinese customer, is they spin the product so quickly, you don't have a 6-year lifetime anymore. It's a 2-3-year lifetime.
Got it. I wanted to come back to the software opportunity for a second, because you, you mentioned $40-$400 ASPs from the low end to high end of CV3. What, what could the software content to Ambarella be if, if you get, you know, some of your modules to the entire full stack, selected by a customer?
I think for example, for the level 4 commercial vehicles, people are talking about, you know, some really high numbers. But I think for consumer vehicles, like Level 2 Plus, Mobileye is talking about $1500 between hardware and software. That give you expectation, right? If it's half a half, I think $700 for hardware is pretty hard. So I would say the $500 hardware and $1000 software, that's what Mobileye charging. I think they are the market leader, so they can charge higher, but I think that set up the range. Maybe everybody else can have the cost lower than that, but that's the range that we're looking at.
But it's still significant.
It's significant.
Yeah. Okay. Last on the auto side, you, you've got a collaboration with CARIAD, VW Software. I don't know if it's still wholly owned by, by VW or if it's been spun off, but, but they're developing the software for the VW family of vehicles, and I think you've announced a partnership with CARIAD. Can you just give us an update on that partnership and when that might start to generate revenue?
We didn't announce CARIAD partnership publicly.
Okay.
I think, I don't think that's we have comments about CARIAD, particularly in any public announcement. I just want to clarify.
Okay.
However, working with any OE and software group is important for us, and working with them, there are two ways. One is that we support them to port software to our hardware. That's default. But the other one is, more importantly, how we can help them to move some module of software into their software stack and speed up the development. I think that's more critical. For example, that joint development with Continental, they contribute a portion of the software, we put another part of software. It's because both company has their strengths and expertise, and we think that we know video better than most company, and our autonomous radar or centralized radar can be a asset for the customer also. So if people want to use that, we are definitely happy to help them to use, to find a way to combine the software.
To have best-of-breed modules in that software stack.
Exactly.
Got it. Okay. I wanted to move to the edge inferencing. You announced the N1 SoC, as you mentioned at CES. You also have the CV72 that I think can be used for edge inferencing. Maybe talk about some of the applications for N1. Where are you seeing sort of customer interest? I know you, you've got the Llama 2 models-
Right
That you're demoing, running on N1, but talk about some of the applications and customer traction.
Right. So first of all, we are not in the cloud, and we don't do training. So that already exclude the majority of business today. But however, we also believe that inference will be bigger opportunity than for training sometime down the road. And, today, we focus on edge and inference only, so that's we need to qualify it first. But even with that, demoing a very low power, low latency, N1 type of solution that can run 17, so sorry, 13 billion. In fact, we just demoed 34 billion in our lab running on N1. That just give plenty flexibility to our customer who are interested in edge AI. Just talk about the current existing customer, like IoT or automotive, all of them are thinking about how to use how LLM going to impact their roadmap.
And I believe all of them will come with an idea, well, helping them to implement the LLM into the software stack. For example, I just hope on the demo that we can have a video camera input to Llama, and the Llama start outputting the description of what they see on the video. That is a demo that nobody can show in the past. And we can see that when you think through the potential application security camera, that's just a strong example. You know, say, a campus has 1,000 cameras. Today, the only thing they can do is when some event happen, they go back to the server, retrieve the video, and try to figure out what happened.
With Llama, we demoed in at CES, you can imagine that for each camera, you sample one picture and feed to the Llama, and so you have 1000 picture into Llama or Llama's application per second. And we can generate report from those 1000 camera every second now. If you accumulate those report on each camera, and you start trying to figure out based on use, doing some existing ChatGPT or search engine, you can summarize what you see on each camera continuously. I think that add tremendous value to people who want to monitor what happens in this campus, right? Just one example, how about traffic cameras, that a city can monitor all the traffic and to find out the traffic flow, you know, identifying, you know, the car being stolen.
Many, many ideas that it can do because now things has to be done by people now can be replaced by Llama. I think that just open the door. Also for automotive, I won't say it's right now, but I think sometime down the road, autonomous driving software will be impacted by Llama model. I don't know when, but I definitely think that's direction people must need to start looking at.
I guess, you know, stepping back, your business has been focused on video.
Yeah
And computer vision, camera vision for the, you know, really since you were founded the company. You know, there are a lot of edge inferencing applications that aren't vision-based.
Right.
That I would think that, that N1 may have applications. And so if you look forward 3-5 years, do you see a, you know, potentially meaningful revenue stream that would not be vision-based? And, you know, do you see potentially Ambarella becoming more of an edge inferencing company, with a heritage, clearly in vision, but, but not necessarily, you know, only in vision?
I think what you just described is exactly what we're going after right now. I think video has been the core of the business for the last 20 years. But moving forward, when you go to the edge inference market, I can see that N1 can serve non-video applications. It will take time, but definitely that we are, like I said, we are focusing on application which require more and more inference AI performance, whether that's video or non-video, become less important for us.
How big do you think this could be in, say, 3-5 years? Is this a $100 million opportunity? You know, what would a reasonable target or expectation for investors to have be?
Well, you know, if you look at all the, you know, research firm publishing number on the inference AI, I have to believe that inference AI is a huge opportunity, but we need to look at which application that we can serve, and at right time, we need to publish our own TAM and SAM numbers based on our research. That's something that we'll probably talk about this at that time. But, I think everybody here understand how big the inference AI market can be. It's really about with the size of Ambarella and our technology, where we should focus on, and optimize our solution for. Because I don't believe in general solution will serve all the market on the edge side.
You can have a general solution for training for the cloud, because that's where you really need generalized solution. But when you come to the edge AI, you need to have a specialized, hardware specialized software targeting our application for power reason, for latency reason, for the, you know, the privacy reason. And I think that will open the different opportunity for us.
Got it. Shifting to the IoT camera part of the business, CV5 is sort of the high end of your CV2 family. Can you give us an update on the ramp there, and do you expect CV5 revenue to be meaningful in fiscal 2025 or effectively calendar 2024?
It will be definitely meaningful, and in fact, we expect that, it ramp up in a big scale, for the IoT business. In fact, we have a design win and also people going to production to support it. I think that's one thing I think continue to. It's consistent with our direction, that even for IoT, we start seeing more and more customer willing to build high-end chip, which demands higher AI performance, and I believe that trend will continue. And the CV5 is for this year. We have CV72 coming up as a pipeline, and we definitely going to continue to have a roadmap for that.
Okay, perfect. How about some of the adjacent markets, like access control or robotics? How's the traction in those segments?
We, you know, we continue to see, opportunity in access control and robotics. Particularly robotics, I think now, in the past, you know, robotics is, so, so segmented, and also you have to just like automotive, right? You have to integrate modules of sensors, planning. You have all kind of layers software. Now, with LLM, I have to I truly believe that the only way to have an integrated hardware and software solution for many different, robotics applications, is using LLM as a, as the core of the software. So you don't need to try to integrate module by module. The LLM would be the brain to integrate everything for you, and people, our customers, are gonna focus on writing applications instead of trying to figure out how to do, you know, perception level or planning level.
Those things should be done by a much integrated solution. With that, then I can see a robotic going into a high-volume integrated business, and I believe that it will take several years to be there, but I definitely consider a robotic as a part of Gen AI application that we need to address.
Perfect. Just lastly, on the competitive landscape, in the IoT camera segment, has there been any changes? Is it still mostly Chinese, Taiwanese competitors at the low end and some of the captive providers at the high end?
I don't think that the competitive landscape change at all.
Yeah.
In fact, that also apply to automotive side. I still think that we compete with, you know, Qualcomm, NVIDIA, and Mobileye on the automotive side. And there are some Chinese companies, like Horizon in China. I don't think the competitive landscape changed, but however, you can see that the venture capitals probably invest on 40, 50 different. And maybe more than that, on the Gen AI or automotive type of companies. I think that, considering how much we invest and where we are at, I think that I will compare to them only when they have a real silicon, real software, then we can do a side-by-side comparison. But I don't think that the competitive landscape changed.
Got it. And, have you seen, through the downturn, any of your competitors getting more aggressive with pricing to try to win sockets, or is the pricing environment fairly stable?
Outside, outside China, no. But inside China, you can see that it's pricing become very important.
Mm-hmm.
But that's why we, you know, in those kind of environment, we have to do really compete on the performance, right? So if we continue to have innovation performance and have a, a, advantage, like on the power side and total BOM, and also, you know, better AI performance, that's how we can protect our, our ASPs. If go on the low end side, in China, really, the, the price is the only differentiation on this side.
Got it. Just wanted to kind of conclude with a few financial questions. On the last call, you talked about being, you know, confident that the revenue in fiscal 2025 would grow, and I think you're expecting growth across both businesses. But maybe just talk about your expectations for IoT and automotive growth. You know, can you say which one's growing faster? And then, can you make any projections for the CV revenue? I think it's about 60% of revenue in fiscal 2024. I assume that goes higher. And thoughts on the video processing revenue?
Right.
Does that start to stabilize?
Right. So first of all, I think both auto and IoT will grow. We haven't quantified what's the percentage, but I think percentage growth rate will probably similar because they, you know, IoT is recovered from the inventory problem, and I think that is going to— It's hard for me to predict the growth because even, like I said, different company come back with a, come back with to finish the inventory, problem at different time. But I definitely feel comfortable with the street number they published because it was average. I think that number reflect our belief also. In terms of, you're talking about, automotive?
Oh, I was just saying, yeah, automotive, but it sounds like you're not, you're not giving sort of, relative, you know, which one grows faster.
I think it's probably similar range, yes.
Okay. And CV3, any projections for where that goes as a % of revenue, or will you wait to maybe quantify that on the next call?
I think for CV3, we talk about our first product in production in 2027, and we believe that for this year, the most important focus is to get design wins for CV3 with the three silicon we announced. I think we are focused a lot in U.S. and Europe, and that's definitely something we think for our investors should track our progress on that.
Okay. Just looking at input pricing was something that, you know, a lot of companies face pressure from higher input prices, both, you know, wafer foundry pricing, as well as other inputs. What is your outlook for wafer and input pricing as we head into calendar 2024? Is it starting to stabilize?
I think stabilize. I think I don't expect a huge change on the input price. But also there are some rumors that the wafer price might come down, too. But I think it's just, you know, I just saw a news that TSMC announced that their their high, the 5-nanometer and 7-nanometer capacity will be utilized fully, but much better utilized than before. So I won't expect the price change a lot on the downside. Whether there will be people going to increase the price, I don't see that either. So I think input price stabilize is probably the best way to describe it.
Got it. And then, just thoughts on OpEx in 2024?
First of all, we don't expect to grow a lot. In fact, in fact, with the current street number, we think it's a little growth, but it's just relatively flat compared to last year. So I think overall, I think that's what we're looking at.
Okay, great. And then sort of lastly, how are you feeling about the company's cash position and sort of uses of, of cash? Will you continue to grow it unless there's attractive M&A, or are there other priorities for the, for the cash?
You know, we have roughly $200 million in bank. I think we feel comfortable with that position, and M&A always a priority. But however, I don't think there is really strong need at this point. You know, radar was a strong need for us, and now we are looking at, really, if we want to continue to go through this AI roadmap, whether there's any other software, hardware company that available to us. But with the current price, I don't think that's going to be a current valuation. That should be a high priority for us. But maintaining a healthy cash position, continue winning automotive design win is important, and I also believe with the current stream numbers, we're going to go back to cash positive for probably the end of this year. That gives us the confidence that we can continue on this route.
Great. All right, well, Fermi, we're at the end of the time, so I really appreciate you joining us here at the Needham Growth Conference. Thank you very much.
Thank you, Quinn. Thank you very much.